Behavior Analysis of Parent-Child Interactions from Text

Parent-Child Interaction Therapy (PCIT) is a ther-apeutic approach designed to enhance the interaction between parents and their children. Parent-child interaction quality may be assessed via observations of free-play or structured tasks between parents and young children. Based on these assess-ments, interventionists provide therapy designed to support more optimal parent-child interactions and promote children's developmental outcomes. While conducting these assessments is vital for children's men-tal health development, manual assessments are time-consuming and resource intensive. This work aims to make PCIT accessible by using Artificial Intelligence (AI) to analyze interaction quality, based on linguistic features in an interactive dialogue. We propose a solution to classify the main behavioral classes in the Dyadic Parent-Child Interaction Coding System (DPICS). To the best of our knowledge, our work is the first model that uses a Transformer-based architecture to analyze the emotions and the psychology integral to the parent-child interactions. The proposed model could understand and detect grammatical, syntactic, and emotional characteristics of the language in parent-child interactions. We categorized Natural Language Processing (NLP) strategies in parent-child interaction quality into three categories: deep learning-based, ML-based, and transfer learning-based. The proposed model is followed by a transfer learning strategy that is fine-tuned over a RoBERTa model and considers text in order to produce comparable results without the use of audio. Our results showed that our proposed model can detect behav-ioral aspects of parent-child interaction without the use of further feature engineering or incorporating additional modalities. We achieved a validation accuracy of 90 %, a significant improvement of 11 % compared to the most successful results reported in similar studies, and therapist agreement rates of 80 %.